diff --git a/04_22_evo.py b/04_22_evo.py
index 53f765d61eb4a46a9c364f63fbd42dd23ec32a7e..cc935e8c68a1eae5ac37922d1a30966aeffc0b1d 100644
--- a/04_22_evo.py
+++ b/04_22_evo.py
@@ -167,17 +167,17 @@ child_network_architecture = cn.LeNet(
img_channels=3
)
-# save_dir='./benchmark/pickles/04_22_cf_ln_rssad'
+save_dir='./benchmark/pickles/04_22_cf_ln_rssad'
# # evo
-# run_benchmark(
-# save_file=save_dir+'.pkl',
-# train_dataset=train_dataset,
-# test_dataset=test_dataset,
-# child_network_architecture=child_network_architecture,
-# agent_arch=aal.EvoLearner,
-# config=config,
-# )
+run_benchmark(
+ save_file=save_dir+'.pkl',
+ train_dataset=train_dataset,
+ test_dataset=test_dataset,
+ child_network_architecture=child_network_architecture,
+ agent_arch=aal.EvoLearner,
+ config=config,
+ )
# # rerun_best_policy(
# # agent_pickle=save_dir+'.pkl',
@@ -194,16 +194,16 @@ child_network_architecture = cn.LeNet(
megapol = [(('ShearY', 0.5, 5), ('Posterize', 0.6, 5)), (('Color', 1.0, 9), ('Contrast', 1.0, 9)), (('TranslateX', 0.5, 5), ('Posterize', 0.5, 5)), (('TranslateX', 0.5, 5), ('Posterize', 0.5, 5)), (('Color', 0.5, 5), ('Posterize', 0.5, 5))]
-accs=[]
-for _ in range(10):
- print(f'{_}/{10}')
- temp_agent = aal.evo_learner(**config)
- accs.append(
- temp_agent.test_autoaugment_policy(megapol,
- child_network_architecture,
- train_dataset,
- test_dataset,
- logging=False)
- )
+# accs=[]
+# for _ in range(10):
+# print(f'{_}/{10}')
+# temp_agent = aal.EvoLearner(**config)
+# accs.append(
+# temp_agent._test_autoaugment_policy(megapol,
+# child_network_architecture,
+# train_dataset,
+# test_dataset,
+# logging=False)
+# )
-print("CIPHAR10 accs: ", accs)
+# print("CIPHAR10 accs: ", accs)
diff --git a/autoaug/autoaugment_learners/EvoLearner.py b/autoaug/autoaugment_learners/EvoLearner.py
index c01aafa857a654072d129b8a724552e1adc1a7bc..9601f2373cd566d88fc24d9250c3e14f2fcc1417 100644
--- a/autoaug/autoaugment_learners/EvoLearner.py
+++ b/autoaug/autoaugment_learners/EvoLearner.py
@@ -95,8 +95,8 @@ class EvoLearner(AaLearner):
def __init__(self,
# search space settings
sp_num=5,
- p_bins=11,
- m_bins=10,
+ p_bins=1,
+ m_bins=1,
discrete_p_m=False,
exclude_method=[],
# child network settings
@@ -131,6 +131,7 @@ class EvoLearner(AaLearner):
# sub_num_pol=self.sp_num
# )
self.controller = controller
+
self.num_solutions = num_solutions
self.torch_ga = torchga.TorchGA(model=self.controller, num_solutions=num_solutions)
self.num_parents_mating = num_parents_mating
@@ -144,6 +145,64 @@ class EvoLearner(AaLearner):
assert num_solutions > num_parents_mating, 'Number of solutions must be larger than the number of parents mating!'
+
+ def get_full_policy(self, x):
+ """
+ Generates the full policy (self.num_sub_pol subpolicies). Network architecture requires
+ output size 5 * 2 * (self.fun_num + self.p_bins + self.m_bins)
+
+ Parameters
+ -----------
+ x -> PyTorch tensor
+ Input data for network
+
+ Returns
+ ----------
+ full_policy -> [((String, float, float), (String, float, float)), ...)
+ Full policy consisting of tuples of subpolicies. Each subpolicy consisting of
+ two transformations, with a probability and magnitude float for each
+ """
+ section = self.fun_num + self.p_bins + self.m_bins
+
+ y = self.controller.forward(x)
+ full_policy = []
+ for pol in range(self.sp_num):
+ int_pol = []
+ for _ in range(2):
+ idx_ret = torch.argmax(y[:, (pol * section):(pol*section) + self.fun_num].mean(dim = 0))
+
+ trans, need_mag = self.augmentation_space[idx_ret]
+
+ if self.p_bins == 1:
+ # p_ret = min(1, max(0, (y[:, (pol * section)+self.fun_num:(pol*section)+self.fun_num+self.p_bins].mean(dim = 0).item())))
+ p_ret = torch.sigmoid(y[:, (pol * section)+self.fun_num:(pol*section)+self.fun_num+self.p_bins].mean(dim = 0)).item()
+ else:
+ p_ret = torch.argmax(y[:, (pol * section)+self.fun_num:(pol*section)+self.fun_num+self.p_bins].mean(dim = 0)).item() * 0.1
+
+ p_ret = round(p_ret, 1)
+
+
+ if need_mag:
+ # print("original mag", y[:, (pol * section)+self.fun_num+self.p_bins:((pol+1)*section)].mean(dim = 0))
+ if self.m_bins == 1:
+ # mag = min(9, max(0, (y[:, (pol * section)+self.fun_num+self.p_bins:((pol+1)*section)].mean(dim = 0).item())))
+ mag = torch.sigmoid(y[:, (pol * section)+self.fun_num+self.p_bins:((pol+1)*section)].mean(dim = 0)).item()
+
+ else:
+ print("bit: ", y[:, (pol * section)+self.fun_num+self.p_bins:((pol+1)*section)].mean(dim = 0))
+ print("full: ", y[:, (pol * section)+self.fun_num+self.p_bins:((pol+1)*section)].shape)
+ print("mean: ", torch.argmax(y[:, (pol * section)+self.fun_num+self.p_bins:((pol+1)*section)].mean(dim = 0)))
+ mag = torch.argmax(y[:, (pol * section)+self.fun_num+self.p_bins:((pol+1)*section)].mean(dim = 0)).item()
+ mag = int(mag)
+ else:
+ mag = None
+ int_pol.append((trans, p_ret, mag))
+
+ full_policy.append(tuple(int_pol))
+
+ return full_policy
+
+
def _get_single_policy_cov(self, x, alpha = 0.5):
"""
@@ -155,7 +214,7 @@ class EvoLearner(AaLearner):
x -> PyTorch Tensor
Input data for the AutoAugment network
- alpha -> Float
+ alpha -> float
Proportion for covariance and population matrices
Returns
@@ -227,20 +286,20 @@ class EvoLearner(AaLearner):
Parameters
------------
- return_weights -> Bool
+ return_weights -> bool
Determines if the weight of the GA network should be returned
Returns
------------
If return_weights:
- Network weights -> Dictionary
+ Network weights -> dict
Else:
Solution -> Best GA instance solution
- Solution fitness -> Float
+ Solution fitness -> float
- Solution_idx -> Int
+ Solution_idx -> int
"""
print("learn0")
self.num_generations = iterations
@@ -260,6 +319,22 @@ class EvoLearner(AaLearner):
def _in_pol_dict(self, new_policy):
+ """
+ Checks if a potential subpolicy has already been testing by the agent
+
+ Parameters
+ ------------
+ new_policy -> subpolicy
+
+ Returns
+ ------------
+ if subpolicy has been tested:
+ -> True
+ else:
+ -> False
+
+
+ """
new_policy = new_policy[0]
trans1, trans2 = new_policy[0][0], new_policy[1][0]
new_set = {new_policy[0][1], new_policy[0][2], new_policy[1][1], new_policy[1][2]}
@@ -276,7 +351,7 @@ class EvoLearner(AaLearner):
def _set_up_instance(self, train_dataset, test_dataset, child_network_architecture):
"""
- Initialises GA instance, as well as fitness and _on_generation functions
+ Initialises GA instance, as well as the fitness and 'on generation' functions
"""
@@ -302,20 +377,27 @@ class EvoLearner(AaLearner):
train_dataset.transform = torchvision.transforms.ToTensor()
self.train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=100)
count = 0
+
+ new_pol = True
for idx, (test_x, label_x) in enumerate(self.train_loader):
- print("here idx: ", idx)
count += 1
- sub_pol = self._get_single_policy_cov(test_x)
+ # sub_pol = self._get_single_policy_cov(test_x)
+ sub_pol = self.get_full_policy(test_x)
+ print("subpol: ", sub_pol)
- while self._in_pol_dict(sub_pol):
- sub_pol = self._get_single_policy_cov(test_x)[0]
+ # if self._in_pol_dict(sub_pol):
+ # sub_pol = self._get_single_policy_cov(test_x)[0]
+ # new_pol = False
+ # fit_val = 0
if idx == 0:
break
print("start test")
- fit_val = self._test_autoaugment_policy(sub_pol,child_network_architecture,train_dataset,test_dataset)
+ if new_pol:
+ fit_val = self._test_autoaugment_policy(sub_pol,child_network_architecture,train_dataset,test_dataset)
+ print("fit_val: ", fit_val)
print("end test")
@@ -342,7 +424,7 @@ class EvoLearner(AaLearner):
def _on_generation(ga_instance):
"""
- Prints information of generational fitness
+ Prints information of generation's fitness
Parameters
-------------
diff --git a/autoaug/autoaugment_learners/GenLearner.py b/autoaug/autoaugment_learners/GenLearner.py
index 3f4a40234285369f775331eb5ae5e913845a7bf5..22630e48087986d9881d533e9c4c756081932705 100644
--- a/autoaug/autoaugment_learners/GenLearner.py
+++ b/autoaug/autoaugment_learners/GenLearner.py
@@ -83,7 +83,7 @@ class Genetic_learner(AaLearner):
self.num_offspring = num_offspring
- def gen_random_subpol(self):
+ def _gen_random_subpol(self):
"""
Generates a random subpolicy using the reduced augmentation_space
@@ -111,7 +111,7 @@ class Genetic_learner(AaLearner):
return subpol
- def gen_random_policy(self):
+ def _gen_random_policy(self):
"""
Generates a random policy, consisting of sp_num subpolicies
@@ -125,7 +125,7 @@ class Genetic_learner(AaLearner):
return pol
- def bin_to_subpol(self, subpol_bin):
+ def _bin_to_subpol(self, subpol_bin):
"""
Converts a binary string representation of a subpolicy to a subpolicy
@@ -166,7 +166,7 @@ class Genetic_learner(AaLearner):
return pol
- def subpol_to_bin(self, subpol):
+ def _subpol_to_bin(self, subpol):
"""
Converts a subpolicy to its binary representation
@@ -200,7 +200,7 @@ class Genetic_learner(AaLearner):
return bin_pol
- def choose_parents(self, parents, parents_weights):
+ def _choose_parents(self, parents, parents_weights):
"""
Chooses parents from which the next policy will be generated from
@@ -224,7 +224,7 @@ class Genetic_learner(AaLearner):
return (parent1, parent2)
- def generate_children(self):
+ def _generate_children(self):
"""
Generates children via the random crossover method
@@ -265,18 +265,15 @@ class Genetic_learner(AaLearner):
"""
for idx in range(iterations):
- print("ITERATION: ", idx)
if len(self.history) < self.num_offspring:
policy = [self.gen_random_subpol()]
else:
policy = self.bin_to_subpol(random.choice(self.generate_children()))
- print("Policy: ", policy)
reward = self._test_autoaugment_policy(policy,
child_network_architecture,
train_dataset,
test_dataset)
- print("reward: ", reward)
diff --git a/autoaug/controller_networks/EvoController.py b/autoaug/controller_networks/EvoController.py
index 33a9a606922483f539f783a33dbf18b1bc16d9cb..db0800f6b52815dc2519402338e6e1ebac939e58 100644
--- a/autoaug/controller_networks/EvoController.py
+++ b/autoaug/controller_networks/EvoController.py
@@ -3,7 +3,7 @@ import torch.nn as nn
import math
class EvoController(nn.Module):
- def __init__(self, fun_num=14, p_bins=11, m_bins=10, sub_num_pol=5):
+ def __init__(self, fun_num=14, p_bins=1, m_bins=1, sub_num_pol=5):
self.fun_num = fun_num
self.p_bins = p_bins
self.m_bins = m_bins